Robin Schmucker

I am a fourth year Ph.D. student in the Machine Learning Department at Carnegie Mellon University, where I am very fortunate to be advised by Professor Tom Mitchell.

Previously, I spent three wonderful years at the Karlsruhe Institute of Technology, completing my bachelor's degree in Computer Science. I was a research assistant and part of the TECO research group lead by Professor Michael Beigl. I am grateful for the mentorship of Professor Manuela Veloso who hosted me as a research intern as part of the CLICS scholarship program.

In the industry, I worked as a research intern at AWS where I designed new algorithms and contributed to the AutoGluon project.

Email  /  CV  /  LinkedIn  /  Google Scholar  /  GitHub

My research interests are broadly in the areas of machine learning and optimization. Particularly, I am interested in the theoretical and practical problems that arise when deploying data driven techniques in real world settings such as education and healthcare. Currently, my research focuses on problems related to intelligent tutoring systems including deep learning for student performance predictions and reinforcement learning for personalized curriculum design.

We propose transfer learning techniques that can mitigate the student performance modeling cold-start problem for new courses by leveraging log data from existing courses. Our course-agnostic models enable accurate predictions for future courses when they are first deployed.

We study how to utilize various types of student log data for performance modeling using four recent large-scale datasets. We propose various extensions over earlier methods and define a new state of the art for logistic regression-based performance modeling.

We use a pan-cancer pathway model to identify novel combination therapies by defining multiple treatment optimization problems and solving them by combining CMA-ES with an efficient Hamiltonian Monte-Carlo sampling scheme. We also consider sequential treatment plans.

We propose multiple algorithms that extend ASHA to the multi-objective hyperparameter optimization (HPO) setting. We assess the performance of our methods on various real world tasks related to neural architecture search, algorithmic fairness, and language model optimization.

We introduce a general constrained Bayesian optimization framework to optimize the performance of any ML model while enforcing different fairness constraints. Our approach is competitive with techniques that enforce model-specific constraints and ones that learn fair representations ahead of time.

We propose the first algorithm for the bandit linear optimization problem for tree-form sequential decision making that offers both (i) linear-time iterations (in the size of the decision tree) and (ii) O(√T) cumulative regret in expectation compared to any fixed strategy, at all times T.

We study the suitability of existing multi-objective algorithms for ML hyperparameter optimization and propose a novel multi-fidelity method. We evaluate on multiple fairness-motivated applications and achieve lower wall-clock times when approximating Pareto frontiers.

We propose the first efficient regret minimization algorithm for the bandit linear optimization problem on sequential decision processes and extensive-form games and show that it achieves O(√T) cumulative regret in expectation against any strategy.

We augment Pepper's perception by integrating state-of-the-art vision and speech recognition systems. As we recognize limitations of the individual perceptual modalities, we introduce a multi-modality approach to increase the robustness of human social interaction with the robot. We combine vision, gesture, speech, and input from an onboard tablet, a remote mobile phone, and external microphones.

By introducing Human Activity Recognition approaches to the robotics domain, we aim at creating agents that can detect their own body’s activities. Our activity recognition pipeline can detect unexpected behavior and can be used to extend Pepper’s inbuilt capabilities.

Quantum mechanics (QM) and the theory of gravity are presently not compatible. One question is whether gravity causes decoherence. We propose a method to estimate gravitational decoherence in an experiment that can draw conclusions in any physical theory where the no-signalling principle holds, even if QM needs to be modified.

Teaching Assistant in Spring 2022.

Teaching Assistant in Fall 2020.

Template of this page: link.